But that’s not the end of the recognition process. Far from it, several specialized units contribute to the end result, and the pretrained neural network or word recognition engine is just one of them. The linguistic context analysis plays its role (and it assigns a confidence level to each character and each word) and so does the typographic analysis (again assigning its own weight).
In other words, once the pretrained neural network or word recognition enginehas recognized each character, several other “expert systems” cast additional votes to validate the results and increase the recognition accuracy. These subsystems allow to validate or correct the result; when alternatives are available, they help in determining which alternative should be selected.
Each expert subsystem plays its role during the recognition process, no single argument by itself takes the day: each expert opinion is given a weight, a confidence level and you move forward until a final decision gets taken. (This approach looks very human, doesn’t it — comparing notes, taking all the factors into consideration before a conclusion is reached? That’s where the real strength of neural networks lie...)
These subsystems are (1) the autolearning, (2) the use of typographic rules and (3) the use of linguistics. We’ll explain the role of “autolearning” next and discuss typography and linguistics later on.
Where does this technology come from? — Reaping the benefits of the neural revolution — … Or do word recognition — Each expert casts his vote — Autolearning font shapes — Putting more feet on the street — … With a document revolution thrown in
Home page — Intro — Scanners — Images — History — OCR — Languages — Accuracy — Output — BCR — Pen scanners — Sitemap — Search — Contact – Feedback